AI learns to build Mario levels by watching YouTube

Tying a double Windsor knot. Making the perfect Pavlova. Speaking Spanish. All things you could learn to do with YouTube, but were too busy watching cat videos to bother with. 

It would appear that computer AIs aren't as easily distracted as their meat sack creators, as one system has just learnt how to play and build Super Mario Bros. levels after learning the process by watching YouTube clips. 

A research team from the Georgia Institute of Technology has built an AI system that "learns" the rules of Mario not by watching players hop the Italian plumber toward Bowser's castle, but by assessing the gaming terrain.

The research results describe how, by judging the "positioning between elements on-screen", a level design rule is learnt and understood by the AI. "For example, pipes in the Mario games tend to stick out of the ground, so the system learns this and prevents any pipes from being flush with grassy surfaces. It also prevents 'breaks' by using spatial analysis – e.g. no impossibly long jumps for the hero."

But the AI doesn't just create pretty levels based on knowing how the building blocks best fit together: by assessing where players spend the majority of their time collecting items, the system is able to learn something of the point-based goals of a Mario level.

With this worryingly precise learning acquired, the AI is then able to build its own pipe-filled worlds with a Super Mario Bros level builder. "Our system creates a model or template, and it’s able to produce level sections that have never been seen before, do not appear random and can be traversed by the player," said Mark Riedl, associate professor of Interactive Computing. 

The AI has built over 300 levels so far, with the next stage of research set to see how humans interact with the computer-constructed worlds.

Sure, it's all very fascinating right now - but what happens when this computer AI system is accidentally plugged into our traffic control systems?! What then?! 

Still, while all this computer learning is very impressive, it's still some way off the Mario knowledge of this guy...

[Via: CNet]